VTechWorks staff will be away for the Thanksgiving holiday beginning at noon on Wednesday, November 27, through Friday, November 29. We will resume normal operations on Monday, December 2. Thank you for your patience.

Browsing by Author "Zhu, Yunhui"

Now showing 1 - 14 of 14
  • Thumbnail Image
    Data-driven X-ray Tomographic Imaging and Applications to 4D Material Characterization
    Wu, Ziling (Virginia Tech, 2021-01-05)
    X-ray tomography is an imaging technique to inspect objects' internal structures with externally measured data by X-ray radiation non-destructively. However, there are concerns about X-ray radiation damage and tomographic acquisition speed in real-life applications. Strategies with insufficient measurements, such as measurements with insufficient dosage (low-dose) and measurements with insufficient projection angles (sparse-view), have been proposed to relieve these problems but are generally compromising imaging quality. Such a dilemma inspires the development of advanced tomographic imaging techniques, in particular, deep learning algorithms to improve reconstruction results with insufficient measurements. The overall aim of this thesis is to design efficient and robust data-driven algorithms with the help of prior knowledge from physics insights and measurement models. We first introduce a hierarchical synthesis CNN (HSCNN), which is a knowledge-incorporated data-driven tomographic reconstruction method for sparse-view and low-dose tomography with a split-and-synthesis approach. This proposed learning-based method informs the forward model biases based on data-driven learning but with reduced training data. The learning scheme is robust against sampling bias and aberrations introduced in the forward modeling. High-fidelity X-ray tomographic imaging reconstruction results are obtained with a very sparse number of projection angles for both numerical simulated and physics experiments. Comparison with both conventional non-learning-based algorithms and advanced learning-based approaches shows improved accuracy and reduced training data size. As a result of the split-and-synthesis strategy, the trained network could be transferable to new cases. We then present a deep learning-based enhancement method, HDrec (hybrid-dose reconstruction algorithm), for low-dose tomography reconstruction via a hybrid-dose acquisition strategy composed of textit{extremely sparse-view normal-dose measurements} and textit{full-view low-dose measurements}. The training is applied for each individual sample without the need of transferring the trained models for other samples. Evaluation of two experimental datasets under different hybrid-dose acquisition conditions shows significantly improved structural details and reduced noise levels compared to results with traditional analytical and regularization-based iterative reconstruction methods from uniform acquisitions under the same amount of total dosage. Our proposed approach is also more efficient in terms of single projection denoising and single image reconstruction. In addition, we provide a strategy to distribute dosage smartly with improved reconstruction quality. When the total dosage is limited, the strategy of combining a very few numbers of normal-dose projections and with not-too-low full-view low-dose measurements greatly outperforms the uniform distribution of the dosage throughout all projections. We finally apply the proposed data-driven X-ray tomographic imaging reconstruction techniques, HSCNN and HDrec, to the dynamic damage/defect characterization applications for the cellular materials and binder jetting additive manufacturing. These proposed algorithms improve data acquisition speeds to record internal dynamic structure changes. A quantitative comprehensive framework is proposed to study the dynamic internal behaviors of cellular structure, which contains four modules: (i) In-situ fast synchrotron X-ray tomography, which enables collection of 3D microstructure in a macroscopic volume; (ii) Automated 3D damage features detection to recognize damage behaviors in different scales; (iii) Quantitative 3D structural analysis of the cellular microstructure, by which key morphological descriptors of the structure are extracted and quantified; (iv) Automated multi-scale damage structure analysis, which provides a quantitative understanding of damage behaviors. In terms of binder jetting materials, we show a pathway toward the efficient acquisition of holistic defect information and robust morphological representation through the integration of (i) fast tomography algorithms, (ii) 3D morphological analysis, and (iii) machine learning-based big data analysis. The applications to two different 4D material characterization demonstrate the advantages of these proposed tomographic imaging techniques and provide quantitative insights into the global evolution of damage/defect beyond qualitative human observation.
  • Thumbnail Image
    High Resolution Phase Imaging using Transport of Intensity Equation
    Shanmugavel, Sibi Chakravarthy (Virginia Tech, 2021-06-23)
    Quantitative phase Imaging(QPI) has emerged as a valuable tool for imaging specimens with weak scattering and absorbing abilities such as cells and tissues. It is complementary to fluorescence microscopy, as such, it can be applied to unlabelled specimens without the need for fluorescent tagging. By quantitatively mapping the phase changes induced in the incident light field by the optical path length delays of the specimen, QPI provides objective measurement of the cellular dynamics and enables imaging the specimen with high contrast. Transport of Intensity Equation(TIE) is a powerful computational tool for QPI because of its experimental and computational simplicity. Using TIE, the phase can be quantitatively retrieved from defocused intensity images. However, the resolution of the phase image computed using TIE is limited by the diffraction limit of the imaging system used to capture the intensity images. In this thesis, we have developed a super resolution phase imaging technique by applying the principles of Structured Illumination Microscopy(SIM) to Transport of Intensity phase retrieval. The modulation from the illumination shifts the high frequency components of the phase object into the system pass-band. This enables phase imaging with resolutions exceeding the diffraction limit. The proposed method is experimentally validated using a custom-made upright microscope. Because of its experimental and computational simplicity, the method in this thesis should find application in biomedical laboratories where super resolution phase imaging is required
  • Thumbnail Image
    High strength and damage-tolerance in echinoderm stereom as a natural bicontinuous ceramic cellular solid
    Yang, Ting; Jia, Zian; Wu, Ziling; Chen, Hongshun; Deng, Zhifei; Chen, Liuni; Zhu, Yunhui; Li, Ling (Nature Research, 2022-10-14)
    Due to their low damage tolerance, engineering ceramic foams are often limited to non-structural usages. In this work, we report that stereom, a bioceramic cellular solid (relative density, 0.2–0.4) commonly found in the mineralized skeletal elements of echinoderms (e.g., sea urchin spines), achieves simultaneous high relative strength which approaches the Suquet bound and remarkable energy absorption capability (ca. 17.7 kJ kg⁻¹) through its unique bicontinuous open-cell foam-likemicrostructure. The high strength is due to the ultra-low stress concentrationswithin the stereom during loading, resulted from their defect-free cellular morphologies with near-constant surface mean curvatures and negative Gaussian curvatures. Furthermore, the combination of bending-induced microfracture of branches and subsequent local jamming of fractured fragments facilitated by small throat openings in stereom leads to the progressive formation and growth of damage bands with significant microscopic densification of fragments, and consequently, contributes to stereom’s exceptionally high damage tolerance.
  • Thumbnail Image
    Investigation of the Processing History during Additive Friction Stir Deposition using In-process Monitoring Techniques
    Garcia, David (Virginia Tech, 2021-02-01)
    Additive friction stir deposition (AFSD) is an emerging solid-state metal additive manufacturing technology that uses deformation bonding to create near-net shape 3D components. As a developing technology, a deeper understanding of the processing science is necessary to establish the process-structure relationships and enable improved control of the as-printed microstructure and material properties. AFSD provides a unique opportunity to explore the friction stir fundamentals via direct observation of the material during processing. This work explores the relationship between the processing parameters (e.g., tool rotation rate Ω, tool velocity V, and material feed rate F) and the thermomechanical history of the material by process monitoring of i) the temperature evolution, ii) the force evolution, and iii) the interfacial contact state between the tool and deposited material. Empirical trends are established for the peak temperature with respect to the processing conditions for Cu and Al-Mg-Si, but a key difference is noted in the form of the power law relationship: Ω/V for Cu and Ω2/V for Al-Mg-Si. Similarly, the normal force Fz for both materials correlates to V and inversely with Ω. For Cu both parameters show comparable influence on the normal force, whereas Ω is more impactful than V for Al-Mg-Si. On the other hand, the torque Mz trends for Al-Mg-Si are consistent with the normal force trends, however for Cu there is no direct correlation between the processing parameters and the torque. These distinct relationships and thermomechanical histories are directly linked to the contact states observed during deformation monitoring of the two material systems. In Cu, the interfacial contact between the material and tool head is characterized by a full slipping condition (δ=1). In this case, interfacial friction is the dominant heat generation mechanism and compression is the primary deformation mechanism. In Al-Mg-Si, the interfacial contact is characterized by a partial slipping/sticking condition (0<δ<1), so both interfacial friction and plastic energy dissipation are important mechanisms for heat generation and material deformation. Finally, an investigation into the contact evolution at different processing parameters shows that the fraction of sticking is critically dependent on the processing parameters which has many implications on the thermomechanical processing history.
  • Thumbnail Image
    Material-Specific Computed Tomography for Molecular X-Imaging in Biomedical Research
    Dong, Xu (Virginia Tech, 2019-04-08)
    X-ray Computed Tomography (CT) imaging has been playing a central role in clinical practice since it was invented in 1972. However, the traditional x-ray CT technique fails to distinguish different materials with similar density, especially for biological tissues. The lack of a quantitative imaging representation has constrained the application of CT technique from a broadening application such as personal or precision medicine. Therefore, my major thesis statement is to develop novel material-specific CT imaging techniques for molecular imaging in biological bodies. To achieve the goal, comprehensive studies were conducted to investigate three different techniques: x-ray fluorescence molecular imaging, material identification (specification) from photon counting CT, and photon counting CT data distortion correction approach based on deep learning. X-ray fluorescence molecular imaging (XFMI) has shown great promise as a low-cost molecular imaging modality for clinical and pre-clinical applications with high sensitivity. In this study, the effects of excitation beam spectrum on the molecular sensitivity of XFMI were experimentally investigated, by quantitatively deriving minimum detectable concentration (MDC) under a fixed surface entrance dose of 200 mR at three different excitation beam spectra. The result shows that the MDC can be readily increased by a factor of 5.26 via excitation spectrum optimization. Furthermore, a numerical model was developed and validated by the experimental data (≥0.976). The numerical model can be used to optimize XFMI system configurations to further improve the molecular sensitivity. Findings from this investigation could find applications for in vivo pre-clinical small-animal XFMI in the future. PCCT is an emerging technique that has the ability to distinguish photon energy and generate much richer image data that contains x-ray spectral information compared to conventional CT. In this study, a physics model was developed based on x-ray matter interaction physics to calculate the effective atomic number () and effective electron density () from PCCT image data for material identification. As the validation of the physics model, the and were calculated under various energy conditions for many materials. The relative standard deviations are mostly less than 1% (161 out of 168) shows that the developed model obtains good accuracy and robustness to energy conditions. To study the feasibility of applying the model with PCCT image data for material identification, both PCCT system numerical simulation and physical experiment were conducted. The result shows different materials can be clearly identified in the − map (with relative error ≤8.8%). The model has the value to serve as a material identification scheme for PCCT system for practical use in the future. As PCCT appears to be a significant breakthrough in CT imaging field, there exists severe data distortion problem in PCCT, which greatly limits the application of PCCT in practice. Lately, deep learning (DL) neural network has demonstrated tremendous success in medical imaging field. In this study, a deep learning neural network based PCCT data distortion correction method was proposed. When applying the algorithm to process the test dataset data, the accuracy of the PCCT data can be greatly improved (RMSE improved 73.7%). Compared with traditional data correction approaches such as maximum likelihood, the deep learning approach demonstrate superiority in terms of RMSE, SSIM, PSNR, and most importantly, runtime (4053.21 sec vs. 1.98 sec). The proposed method has the potential to facilitate the PCCT studies and applications in practice.
  • Thumbnail Image
    Mechanical Design of Selected Natural Ceramic Cellular Solids
    Yang, Ting (Virginia Tech, 2021-05-24)
    While the structure and mechanical properties of natural cellular solids such as wood and trabecular bone have been extensively studied in the past, the structural design and underlying deformation mechanisms of natural cellular solids with very high mineral contents (> 90 wt%), which we term as natural ceramic cellular solids, are largely unexplored. Many of these natural ceramic cellular solids, despite their inherent brittle constituent biominerals (e.g., calcite or aragonite), exhibit remarkable mechanical properties, such as high stiffness and damage tolerance. In this thesis, by carefully selecting three biomineralized skeletal models with distinctly different cellular morphologies, including the honeycomb-like structure in cuttlefish bone (or cuttlebone), the stochastic open-cell structure in sea urchin spines, and the periodic open-cell structure in starfish ossicles, I systematically investigate the mechanical design strategies of these natural ceramic cellular solids. The three model systems are cuttlefish Sepia officinalis, sea urchin Heterocentrotus mammillatus, and starfish Protoreaster nodosus, respectively. By investigating the relationship between their mechanical properties and structural characteristics, this thesis reveals some novel structural design strategies for developing lightweight, tough, strong, and stiff ceramic cellular solids. The internal skeleton of S. officinalis, also known as cuttlebone, has a porosity of 93 vol% (constituent material: 90 wt% aragonite), which is a multichambered structure consisting of horizontal septa and thin vertical walls with corrugated cross-sectional profiles. Through systematic ex-situ and synchrotron-based in-situ mechanical measurements and collaborative computational modeling, we reveal that the vertical walls in the cuttlebone exhibit an optimal waviness gradient, which leads to compression-dominant deformation and asymmetric wall fracture, accomplishing both high stiffness (8.4 MN∙m/kg) and high energy absorption (4.4 kJ/kg). Moreover, the distribution of walls reduces stress concentrations within the horizontal septa, facilitating a larger chamber crushing stress and more significant densification. For the stochastic open-cell foam-like structure, also known as stereom (porosity: 60-80 vol%, constituent material: 99 wt% calcite) in H. mammillatus, we first developed a computer vision-based algorithm that allows for quantitative analysis of the cellular network of these structures at both local individual branch and node level as well as the global network level. This open-source algorithm could be used for analyzing both biological and engineering open-cell foams. I further show that the smooth, highly curved branch morphology with near-constant surface curvature in stereom results in low-stress concentration, which further leads to dispersed crack formation upon loading. Combined synchrotron in-situ analysis, electron microscopic analysis, and computational modeling further reveal that the fractured branches are efficiently jammed by the small throat openings within the cellular structure. This further leads to the formation of damage bands with densely packed fracture pieces. The continuous widening of the damage bands through progressive microfracture of branches at the boundaries contributes to the observed high plateau stress during compression, thereby contributing to its high energy absorption (17.7 kJ/kg), which is comparable and even greater than many synthetic metal- and polymer-based foams. Lastly, this thesis leads to the discovery of a unique dual-scale single-crystalline porous lattice structure (porosity: 50 vol%, constituent material: 99 wt% calcite) in the ossicles of P. nodosus. At the atomic level, the ossicle is composed of single-crystal biogenic calcite. At the lattice level, the ossicle's microstructure organizes as a diamond-triply periodic minimal surface (TPMS) structure. Moreover, the crystallographic axes at atomic and lattice levels are aligned, i.e., the c-axis of calcite is aligned with the [111] direction of the diamond-TPMS lattice. This single crystallinity co-alignment at two levels mitigates the compliance of calcite in the c-axis direction by utilizing the stiff <111> direction of the diamond-TPMS lattice. Furthermore, 3D in-situ mechanical characterizations reveal that the presence of crystal defects such as 60° and screw dislocations at the lattice level suppresses slip-like fracture along the {111} planes of the calcitic diamond-TPMS lattice upon loading, achieving an enhanced energy absorption capability. Even though the skeleton of echinoderm is made up of single-crystal calcite, the structure fractures in a conchoidal manner rather than along the clipping plane, which can continuously fracture the fragments into small pieces and enhance energy dissipation.
  • Thumbnail Image
    Microstructure Representation and Prediction via Convolutional Neural Network-Based Texture Representation and Synthesis, Towards Process Structure Linkage
    Han, Yi (Virginia Tech, 2021-05-19)
    Metal additive manufacturing (AM) provides a platform for microstructure optimization via process control, the ability to model the evolution of microstructures from changes in processing condition or even predict the microstructures from given processing condition would greatly reduce the time frame and the cost of the optimization process. In 1, we present a deep learning framework to quantitatively analyze the microstructural variations of metals fabricated by AM under different processing conditions. We also demonstrate the capability of predicting new microstructures from the representation with deep learning and we can explore the physical insights of the implicitly expressed microstructure representations. We validate our framework using samples fabricated by a solid-state AM technology, additive friction stir deposition, which typically results in equiaxed microstructures. In 2, we further improve and generalize the generating framework, a set of metrics is used to quantitatively analyze the effectiveness of the generation by comparing the microstructure characteristics between the generated samples and the originals. We also take advantage of image processing techniques to aid the calculation of metrics that require grain segmentation.
  • Thumbnail Image
    Neural Networks For Phase Demodulation In Optical Interferometry
    Black, Jacob A. (Virginia Tech, 2019)
    Neural Networks (NNs) (or 'deep' neural networks (DNNs)) have found great success in many applications across all fields of engineering, and in particular have found recent success in the field of Photonics. In this work we discuss the application of NNs to optical interferometry for the purpose of quantitative phase imaging (QPI). We show that NNs are capable of quantifying the optical pathlength difference in an interferogram with sensitivities that achieve the fundamental limit given by the Cramér-Rao bound (CRB). As an application, we consider a particular QPI technique known as wavelength shifting interferometry (WSI) which obtains the OPL by acquiring multiple interferograms at different, evenly spaced wavenumbers. Traditional phase demodulation algorithms for WSI fail to reach the theoretical OPL sensitivity limit set by the CRB. We have designed NNs which are capable of achieving this bound across a wide range of OPL differences. The NNs are trained on simulated data, and then applied to experimental data. In both simulation and experiment, the NNs outperform the existing analytical demodulation techniques and provide highly sensitive signal demodulation in cases where the analytical approach fails. Thus, NNs provide better performance and more flexibility in the design and use of a WSI system. We expect that the techniques developed in this work can be extended to other two-beam interference based QPI system.
  • Thumbnail Image
    Non-linearity and Dispersion Effects in Tissue Impedance during Application of High Frequency Electroporation-Inducing Pulsed Electric Fields
    Bhonsle, Suyashree P. (Virginia Tech, 2018-01-27)
    Since its conception in 2005, irreversible electroporation (IRE), a non-thermal tumor ablation modality, was investigated for safety and efficacy in clinical applications concerning different organs. IRE utilizes high voltage (~3kV), short duration (~100us) pulses to create transient nanoscale defects in the plasma membrane to cause cell death due to irreversible defects, osmotic imbalances and ATP loss. More recently, high-frequency irreversible electroporation (H-FIRE), which employs narrow bipolar pulses (~0.5-10us) delivered in bursts (on time ~100us), was invented to provide benefits such as the mitigation of intense muscle contractions associated with IRE-based treatments. Furthermore, H-FIRE exhibits the potential to improve lesion predictability in homogeneous and heterogeneous tissue masses. Therapeutic IRE and H-FIRE utilize source and sink electrodes inserted into or around the tumor to deliver the treatment. Prediction of the ablation size, for a set of parameters, can be achieved by the use of pre-treatment planning algorithms that calculate the induced electric field distribution in the target tissue. An electric field above a certain threshold induces cell death and parameters are tuned to ensure complete tumor coverage while sparing the nearby healthy tissue. IRE studies have shown that the underlying field is influenced by the increase in tissue conductivity due to enhanced membrane permeability, and treatment outcome can be improved when this nonlinearity is accounted for in numerical models. Since IRE pulses far exceed the time constant of the cell (~1us), the tissue response can be treated as essentially DC a static approximation can be used to predict the field distribution. Alternately, as H-FIRE pulses are on the order of the time constant of the membrane, the tissue response can no longer be treated as DC. The complexity of the H-FIRE-induced field distribution is further enhanced due to the dispersion and non-linearity in biological tissue impedance during treatment. In this dissertation, we have studied the electromagnetic fields induced in tissue during H-FIRE using several experimental and modeling techniques. In addition, we have characterized the nonlinearity and dispersion in tissue impedance during H-FIRE treatments and proposed simpler methods to predict the field distribution to enable easier translation to the clinic.
  • Thumbnail Image
    Personalized Voice Activated Grasping System for a Robotic Exoskeleton Glove
    Guo, Yunfei (Virginia Tech, 2021-01-05)
    Controlling an exoskeleton glove with a highly efficient human-machine interface (HMI), while accurately applying force to each joint remains a hot topic. This paper proposes a fast, secure, accurate, and portable solution to control an exoskeleton glove. This state of the art solution includes both hardware and software components. The exoskeleton glove uses a modified serial elastic actuator (SEA) to achieve accurate force sensing. A portable electronic system is designed based on the SEA to allow force measurement, force application, slip detection, cloud computing, and a power supply to provide over 2 hours of continuous usage. A voice-control-based HMI referred to as the integrated trigger-word configurable voice activation and speaker verification system (CVASV), is integrated into a robotic exoskeleton glove to perform high-level control. The CVASV HMI is designed for embedded systems with limited computing power to perform voice-activation and voice-verification simultaneously. The system uses MobileNet as the feature extractor to reduce computational cost. The HMI is tuned to allow better performance in grasping daily objects. This study focuses on applying the CVASV HMI to the exoskeleton glove to perform a stable grasp with force-control and slip-detection using SEA based exoskeleton glove. This research found that using MobileNet as the speaker verification neural network can increase the speed of processing while maintaining similar verification accuracy.
  • Thumbnail Image
    Physics-informed neural network for phase imaging based on transport of intensity equation
    Wu, Xiaofeng; Wu, Ziling; Shanmugavel, Sibi Chakravarthy; Yu, Hang Z.; Zhu, Yunhui (Optica Publishing Group, 2022-11)
    Non-interferometric quantitative phase imaging based on Transport of Intensity Equation (T1E) has been widely used in bio-medical imaging. However, analytic TIE phase retrieval is prone to low-spatial frequency noise amplification, which is caused by the iliposedness of inversion at the origin of the spectrum. There are also retrieval ambiguities resulting from the lack of sensitivity to the curl component of the Poynting vector occurring with strong absorption. Here, we establish a physics-informed neural network (PINN) to address these issues, by integrating the forward and inverse physics models into a cascaded deep neural network. We demonstrate that the proposed PINN is efficiently trained using a small set of sample data, enabling the conversion of noise-corrupted 2-shot TIE phase retrievals to high quality phase images under partially coherent LED illumination. The efficacy of the proposed approach is demonstrated by both simulation using a standard image database and experiment using human buccal epitehlial cells. In particular, high image quality (SSIM = 0.919) is achieved experimentally using a reduced size of labeled data (140 image pairs). We discuss the robustness of the proposed approach against insufficient training data, and demonstrate that the parallel architecture of PINN is efficient for transfer learning.
  • Thumbnail Image
    Primary/Soft Biometrics: Performance Evaluation and Novel Real-Time Classifiers
    Alorf, Abdulaziz Abdullah (Virginia Tech, 2020-02-19)
    The relevance of faces in our daily lives is indisputable. We learn to recognize faces as newborns, and faces play a major role in interpersonal communication. The spectrum of computer vision research about face analysis includes, but is not limited to, face detection and facial attribute classification, which are the focus of this dissertation. The face is a primary biometric because by itself revels the subject's identity, while facial attributes (such as hair color and eye state) are soft biometrics because by themselves they do not reveal the subject's identity. In this dissertation, we proposed a real-time model for classifying 40 facial attributes, which preprocesses faces and then extracts 7 types of classical and deep features. These features were fused together to train 3 different classifiers. Our proposed model yielded 91.93% on the average accuracy outperforming 7 state-of-the-art models. We also developed a real-time model for classifying the states of human eyes and mouth (open/closed), and the presence/absence of eyeglasses in the wild. Our method begins by preprocessing a face by cropping the regions of interest (ROIs), and then describing them using RootSIFT features. These features were used to train a nonlinear support vector machine for each attribute. Our eye-state classifier achieved the top performance, while our mouth-state and glasses classifiers were tied as the top performers with deep learning classifiers. We also introduced a new facial attribute related to Middle Eastern headwear (called igal) along with its detector. Our proposed idea was to detect the igal using a linear multiscale SVM classifier with a HOG descriptor. Thereafter, false positives were discarded using dense SIFT filtering, bag-of-visual-words decomposition, and nonlinear SVM classification. Due to the similarity in real-life applications, we compared the igal detector with state-of-the-art face detectors, where the igal detector significantly outperformed the face detectors with the lowest false positives. We also fused the igal detector with a face detector to improve the detection performance. Face detection is the first process in any facial attribute classification pipeline. As a result, we reported a novel study that evaluates the robustness of current face detectors based on: (1) diffraction blur, (2) image scale, and (3) the IoU classification threshold. This study would enable users to pick the robust face detector for their intended applications.
  • Thumbnail Image
    Regularization, Uncertainty Estimation and Out of Distribution Detection in Convolutional Neural Networks
    Krothapalli, Ujwal K. (Virginia Tech, 2020-09-11)
    Classification is an important task in the field of machine learning and when classifiers are trained on images, a variety of problems can surface during inference. 1) Recent trends of using convolutional neural networks (CNNs) for various machine learning tasks has borne many successes and CNNs are surprisingly expressive in their learning ability due to a large number of parameters and numerous stacked layers in the CNNs. This increased model complexity also increases the risk of overfitting to the training data. Increasing the size of the training data using synthetic or artificial means (data augmentation) helps CNNs learn better by reducing the amount of over-fitting and producing a regularization effect to improve generalization of the learned model. 2) CNNs have proven to be very good classifiers and generally localize objects well; however, the loss functions typically used to train classification CNNs do not penalize inability to localize an object, nor do they take into account an object's relative size in the given image when producing confidence measures. 3) Convolutional neural networks always output in the space of the learnt classes with high confidence while predicting the class of a given image regardless of what the image consists of. For example an ImageNet-1K trained CNN can not say if the given image has no objects that it was trained on if it is provided with an image of a dinosaur (not an ImageNet category) or if the image has the main object cut out of it (context only). We approach these three different problems using bounding box information and learning to produce high entropy predictions on out of distribution classes. To address the first problem, we propose a novel regularization method called CopyPaste. The idea behind our approach is that images from the same class share similar context and can be 'mixed' together without affecting the labels. We use bounding box annotations that are available for a subset of ImageNet images. We consistently outperform the standard baseline and explore the idea of combining our approach with other recent regularization methods as well. We show consistent performance gains on PASCAL VOC07, MS-COCO and ImageNet datasets. For the second problem we employ objectness measures to learn meaningful CNN predictions. Objectness is a measure of likelihood of an object from any class being present in a given image. We present a novel approach to object localization that combines the ideas of objectness and label smoothing during training. Unlike previous methods, we compute a smoothing factor that is adaptive based on relative object size within an image. We present extensive results using ImageNet and OpenImages to demonstrate that CNNs trained using adaptive label smoothing are much less likely to be overconfident in their predictions, as compared to CNNs trained using hard targets. We train CNNs using objectness computed from bounding box annotations that are available for the ImageNet dataset and the OpenImages dataset. We perform extensive experiments with the aim of improving the ability of a classification CNN to learn better localizable features and show object detection performance improvements, calibration and classification performance on standard datasets. We also show qualitative results using class activation maps to illustrate the improvements. Lastly, we extend the second approach to train CNNs with images belonging to out of distribution and context using a uniform distribution of probability over the set of target classes for such images. This is a novel way to use uniform smooth labels as it allows the model to learn better confidence bounds. We sample 1000 classes (mutually exclusive to the 1000 classes in ImageNet-1K) from the larger ImageNet dataset comprising about 22K classes. We compare our approach with standard baselines and provide entropy and confidence plots for in distribution and out of distribution validation sets.
  • Thumbnail Image
    Surface Modification of Multimaterial Multifunctional Fibers Enabling Biosensing Applications
    Lopez Marcano, Ana Graciela (Virginia Tech, 2018-06-27)
    During the last decades, the continuing need for faster and smaller sensors has indeed triggered the rapid growth of more sophisticated technologies. This has led to the development of new optical-based sensors, able to detect and measure different phenomena using light. Furthermore, material processing technologies and micro fabrication methods have exponentially advanced, allowing engineers and scientists to develop new and more complex sensors on optical fibers platforms; specifically attractive for life science and biomedical research. All these substantial developments have brought biosensors to a point where multifunctionality is needed, this has led to envision the "Lab-on-Fiber" concept. Which promotes the integration of different sensing components into a single platform, an optical fiber. In this work, an integrated system with non-conventional polymer optical fibers and their further surface modification has been developed. With these different approaches, electrodes, hollow channels and plasmonic nanostructures can be incorporated into a single optical fiber-based sensor, allowing for both electrical and optical sensing with the capabilities of tuning and signal enhancement thanks to the metallic nanostructures. Different fiber substrates can be designed and modified in order to satisfy multiple requirements for a wide variety of applications.